ethanolivertroy/hipaa-compliance-training

HIPAA Compliance Training Dataset

Dataset Description

The first comprehensive HIPAA compliance training dataset for LLM fine-tuning, covering the Security Rule, Privacy Rule, Breach Notification Rule, and implementation guidance from NIST and FDA.

Dataset Summary

• Total Examples: 1,287 (1,029 train / 258 validation)
• Source Documents: 9 federal publications (~5.6 MB extracted content)
• Format: JSONL with chat-formatted messages
• License: CC0-1.0 (Public Domain)
• Coverage: Security Rule, Privacy Rule, Breach Notification, NIST implementation, FDA medical device guidance

Why This Dataset?

Healthcare organizations need practical HIPAA compliance guidance, but most LLMs lack deep knowledge of:

• The distinction between required vs addressable implementation specifications
• Covered entity vs business associate obligations
• Administrative, physical, and technical safeguards
• Breach notification thresholds and procedures
• Practical implementation of NIST SP 800-66 recommendations

This dataset fills that gap with 1,287 high-quality Q&A pairs derived from authoritative federal sources.

Dataset Structure

Data Fields

• messages: List of chat messages (system, user, assistant)
• metadata: Source attribution and HIPAA-specific classification
  • source: Source document title
  • cfr_section: CFR section number (e.g., §164.308) if applicable
  • safeguard_type: administrative | physical | technical | unknown
  • implementation_spec: required | addressable | both
  • entity_type: covered_entity | business_associate | both
  • type: Question type (cfr_section, safeguard, implementation, breach_notification, etc.)
  • nist_control: NIST control ID (e.g., AC-1) if from SP 800-66

Example

{
  "messages": [
    {
      "role": "system",
      "content": "You are a HIPAA compliance expert with deep knowledge of the Security Rule, Privacy Rule, and Breach Notification Rule. You provide accurate, practical guidance on HIPAA requirements, implementation specifications, and compliance procedures. You cite specific CFR sections and distinguish between covered entity and business associate obligations."
    },
    {
      "role": "user",
      "content": "What's the difference between required and addressable implementation specifications?"
    },
    {
      "role": "assistant",
      "content": "According to HITECH Omnibus Final Rule (2013):\n\nUnder the HIPAA Security Rule, implementation specifications are either Required or Addressable. Required specifications must be implemented as written. Addressable specifications require a more flexible approach: covered entities and business associates must assess whether each addressable specification is a reasonable and appropriate safeguard in their environment. If reasonable and appropriate, they must implement the specification as written. If not, they must document why the specification is not reasonable and appropriate, and implement an equivalent alternative measure if reasonable and appropriate, or document why no alternative measure is reasonable and appropriate and accept the residual risk. This risk-based, scalable approach allows entities of different sizes and complexities to implement appropriate security measures."
    }
  ],
  "metadata": {
    "source": "HITECH Omnibus Final Rule (2013)",
    "implementation_spec": "both",
    "safeguard_types": ["administrative", "physical", "technical"],
    "type": "implementation"
  }
}

Source Documents

Federal Register Preambles (5 documents)

These provide the rationale and implementation guidance behind HIPAA regulations:

1. Security Rule Final Rule (2003) - 344 KB

   • Security standards rationale
   • Administrative, physical, and technical safeguards
   • Required vs addressable specifications

2. Privacy Rule Final Rule (2000) - 2.6 MB

   • Privacy standards and uses/disclosures
   • Individual rights (access, amendment, accounting)
   • Minimum necessary standard

3. Privacy Rule Modifications (2002) - 672 KB

   • Clarifications and corrections
   • Business associate provisions
   • De-identification standards

4. HITECH Omnibus Final Rule (2013) - 1.0 MB

   • HITECH Act modifications
   • Business associate direct liability
   • Breach notification updates
   • Enhanced enforcement

5. Breach Notification Rule (2009) - 228 KB

   • Breach definition and risk assessment
   • Notification requirements (HHS, individuals, media)
   • 500+ individual threshold

NIST Implementation Guidance (1 document)

6. NIST SP 800-66 Rev. 2: Implementing the HIPAA Security Rule (2024) - 497 KB
   • Step-by-step implementation guidance
   • Control mapping (NIST 800-53 → HIPAA)
   • Risk analysis methodology
   • Practical examples and recommendations

FDA Medical Device Cybersecurity (2 documents)

7. FDA Cybersecurity in Medical Devices: Quality System Considerations (2023) - 211 KB

   • Device security requirements
   • Cybersecurity bill of materials (SBOM)
   • Premarket submission requirements

8. FDA Postmarket Management of Cybersecurity in Medical Devices (2016) - 84 KB

   • Vulnerability management
   • Patching and updates
   • Incident response for devices

Documents Excluded (Extraction Failures)

The following 4 documents failed to extract properly and are not included:

• 45 CFR Part 160, 162, 164 (data format errors - extracted HTML navigation only)
• FDA Content of Premarket Submissions Draft Guidance (2022) (extraction failed - 71 bytes)

These will be added in v1.1 using alternative extraction methods.

Q&A Generation Strategies

Examples were generated using 7 complementary strategies:

1. CFR Section-Based Q&A

Extract specific regulatory requirements by CFR section number.

Example: "What does §164.308(a)(1)(i) require?" → Security Management Process details

2. Safeguard-Based Q&A

Focus on administrative, physical, and technical safeguards.

Example: "What are the technical safeguards under HIPAA?" → Access control, audit controls, integrity, transmission security

3. Implementation Specification Q&A

Clarify required vs addressable distinctions.

Example: "How do I determine if an addressable specification must be implemented?" → Risk assessment process

4. Entity Type Q&A

Distinguish covered entity vs business associate obligations.

Example: "What are business associate obligations under HIPAA?" → Direct liability, security, privacy, breach notification

5. Breach Notification Q&A

Explain breach determination and notification procedures.

Example: "When must a HIPAA breach be reported to HHS?" → 60-day timeline, 500+ threshold, risk assessment

6. NIST Control Q&A

Map NIST 800-53 controls to HIPAA requirements (from SP 800-66).

Example: "How does NIST SP 800-66 recommend implementing control AC-2?" → Account management for HIPAA

7. Semantic Chunking

General content preservation with context.

Example: "What does [document] say about HIPAA compliance?" → Contextual excerpts

Use Cases

Healthcare Providers

• Internal HIPAA chatbot: Answer staff questions about compliance
• Training assistant: Onboard new employees on HIPAA requirements
• Policy generation: Draft security policies based on HIPAA guidance

Health Tech Companies

• Compliance review: Check product features against HIPAA requirements
• Risk assessment: Identify gaps in HIPAA implementation
• BAA review: Understand business associate obligations

Medical Device Manufacturers

• Device security: Implement FDA cybersecurity requirements
• SBOM generation: Create cybersecurity bills of materials
• Vulnerability management: Follow FDA postmarket guidance

HIPAA Compliance Officers

• Breach determination: Assess if incidents require notification
• Audit preparation: Review requirements before assessments
• Implementation planning: Prioritize required vs addressable specs

Legal & Consulting

• Client advisory: Provide accurate HIPAA interpretations
• Incident response: Guide breach notification decisions
• Contract review: Verify BAA compliance

Dataset Statistics

Source Documents:         9 (5 preambles, 1 NIST, 2 FDA, 1 report)
Total Examples:           1,287
Training Examples:        1,029 (80%)
Validation Examples:      258 (20%)
Avg Example Length:       ~400 tokens (estimated)

By Safeguard Type:
  Administrative:         8 documents
  Physical:               8 documents
  Technical:              8 documents
  Unknown:                3 documents

By Entity Type:
  Both (CE & BA):         9 documents
  Covered Entity:         2 documents

Breach-Related:           8 documents

By Implementation Spec:
  Both (Required & Addressable): 5 documents
  Required only:          6 documents

Model Training

Recommended Configuration

This dataset is designed for LoRA fine-tuning of 7B parameter models:

base_model: mlx-community/Qwen2.5-Coder-7B-Instruct-4bit
lora_rank: 16
lora_alpha: 32
num_layers: 16
batch_size: 4
learning_rate: 1e-5
num_iterations: 500-1000

Expected Training Time

• 7B model on M4 Max: ~45-90 minutes (500-1000 iterations)
• Memory required: 15-20 GB
• Tokens/sec: 130-160

Training Command

python -m mlx_lm lora \
  --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
  --train \
  --data data/hipaa/training \
  --iters 1000 \
  --batch-size 4 \
  --lora-layers 16 \
  --adapter-path models/hipaa-expert-lora

Quick Start

Load Dataset

import jsonlines

# Load training data
with jsonlines.open('train.jsonl') as reader:
    train_data = list(reader)

print(f"Training examples: {len(train_data)}")
print(f"First example: {train_data[0]}")

Filter by Metadata

# Find all breach notification examples
breach_examples = [
    ex for ex in train_data
    if ex['metadata'].get('type') == 'breach_notification'
]

# Find technical safeguard examples
technical_examples = [
    ex for ex in train_data
    if 'technical' in ex['metadata'].get('safeguard_types', [])
]

# Find business associate-specific content
ba_examples = [
    ex for ex in train_data
    if ex['metadata'].get('entity_type') == 'business_associate'
]

Limitations

Coverage Gaps

1. Missing CFR Regulations: The actual 45 CFR Part 160, 162, 164 text failed to extract (data format errors). Dataset relies on preambles and guidance, which provide implementation context but not the bare regulatory text.

2. Preamble-Heavy: 5 of 9 documents are Federal Register preambles. These provide rationale and interpretation but may be verbose compared to concise regulatory text.

3. No OCR Guidance: HHS Office for Civil Rights FAQs and guidance documents are not included (would require web scraping).

4. Limited Medical Device Coverage: Only 2 FDA guidance documents vs comprehensive device security library.

5. No State Law Extensions: California CMIA, Massachusetts 201 CMR 17.00, etc. not included.

Model Limitations

• Not Legal Advice: Outputs should not be considered legal advice
• May Hallucinate: Model may generate plausible but incorrect CFR sections
• Context Window: Limited to base model's context (typically 4K-8K tokens)
• No Real-Time Updates: Trained on documents up to 2024; regulations may change

Recommended Mitigations

1. RAG Integration: Use with vector database of actual HIPAA regulations
2. Citation Verification: Always verify CFR sections in official sources
3. Human Review: Have compliance officers review model outputs
4. Regular Updates: Retrain annually as regulations evolve

Roadmap

v1.1 (Planned)

• Re-extract CFR regulations using alternative methods (PDF repair tools)
• Add OCR guidance via web scraping HHS.gov/hipaa
• Expand FDA coverage with additional medical device guidance
• Target: +500-1,000 examples

v1.2 (Planned)

• State-specific HIPAA extensions (CA CMIA, MA 201 CMR 17.00, NY SHIELD Act)
• CMS security standards
• FTC Health Breach Notification Rule
• Target: +300-500 examples

v2.0 (Future)

• Full OCR guidance library (200-400 FAQs)
• Settlement agreement case studies
• Enforcement examples
• Synthetic scenario generation
• Target: +1,000-2,000 examples

Citation

If you use this dataset, please cite:

@dataset{hipaa_compliance_2025,
  title={HIPAA Compliance Training Dataset},
  author={Troy, Ethan Oliver},
  year={2025},
  publisher={HuggingFace},
  note={Derived from federal HIPAA regulations and guidance (Public Domain)}
}

License

CC0-1.0 (Public Domain) - All source documents are U.S. Federal Government works not subject to copyright protection under 17 U.S.C. §105.

You are free to:

• Use commercially
• Modify and distribute
• Create derivative works
• No attribution required (though appreciated!)

Acknowledgments

This dataset is built from publications by:

• HHS Office for Civil Rights (OCR) - HIPAA regulations and guidance
• National Institute of Standards and Technology (NIST) - SP 800-66 implementation guide
• U.S. Food and Drug Administration (FDA) - Medical device cybersecurity guidance
• Federal Register - Official regulatory preambles

Contact

For questions, issues, or contributions:

• GitHub: https://github.com/ethanolivertroy/nist-tuned-model
• HuggingFace: https://huggingface.co/ethanolivertroy
• Dataset Issues: Open an issue on GitHub

Version History

• v1.0 (January 2025): Initial release
  • 1,287 examples from 9 federal documents
  • Coverage: Security Rule, Privacy Rule, Breach Notification, NIST SP 800-66, FDA guidance
  • Known limitation: CFR regulations failed to extract (data format errors)

---

Last Updated: January 2025